Bearing capacity prediction of shallow foundation on sandy soils: a comparative study of analytical, FEM, and machine learning approaches

In this study, we compared the Ultimate Bearing Capacity (UBC) of shallow foundations on sandy soils that were predicted using Analytical, Finite Element Modeling (FEM), and Machine Learning (ML) approaches for predicting the Ultimate Bearing Capacity (UBC) of shallow foundations on sandy soils. For the first of its type, we presented a novel Python-based pipeline that enables rapid and precise estimation of the UBC for shallow foundations, surpassing traditional methods by providing superior speed and accuracy. The proposed models consider the foundations’ geometry and soil properties as input parameters. We created, trained, and tested nineteen ML models using the Pycaret library in the Google Colab environment. Furthermore, we conducted a comparative analysis of twelve new datasets derived from the training process. Our objective was to estimate the UBC values using three established techniques: (a) the widely recognized Terzaghi method, (b) the advanced three-dimensional FEM software (using OptumG3 software), and the ML-based method. Based on the ML results, we found that Gradient Boosting Regression (gbr), AdaBoost Regression (ada), Random Forest Regression (rf), and Extra Tree Regression (et) were the most effective models for estimating UBC. The gbr model exhibited the highest UBC prediction performance, attaining an R2 value of 1 on the training set, an R2 value of 0.937 on the test set, and an RMSE of 1.171 kPa. Using sensitivity analysis results, we demonstrated that the internal friction angle of the soil is the most significant input variable for estimating UBC, closely followed by the depth of the footing. The comparative results revealed that the well-known Terzaghi method and FEM modeling underestimate the UBC. The proposed user-friendly pipeline would be a valuable tool for geotechnical engineers to predict UBC values, providing a larger dataset in future research that can be trained and tested for the model to enhance reliability further.